adonis vs. permanova

[baoanxh2006]

Dear Qiime Users,

I am using adonis and permanova to find the role of environmental conditions (redox and treatment) on the variation of microbial community.

I noted this thread: https://groups.google.com/forum/#!newtopic/qiime-forum/qiime-forum/qyDJAdECBnc.

However, I have not been able to find out my problems.

With treatment (control vs. osa), I got the same pseudo-F statistic by both methods:

Call:

adonis(formula = as.dist(qiime.data$distmat) ~ qiime.data$map[[opts$category]],      permutations = opts$num_permutations) 

Permutation: free

Number of permutations: 999

Terms added sequentially (first to last)

                                Df SumsOfSqs MeanSqs F.Model      R2 Pr(>F)   

qiime.data$map[[opts$category]]  1   0.39678 0.39678  2.7742 0.17587  0.003 **

Residuals                       13   1.85931 0.14302         0.82413          

Total                           14   2.25609                 1.00000          

---

Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

method name PERMANOVA

test statistic name pseudo-F

sample size 15

number of groups 2

test statistic 2.7741947452989191

p-value 0.002

number of permutations 999

However, with redox condition (200 vs. 250 vs. -100 vs. -400), two methods produced complete different pseudo-F values:

Call:

adonis(formula = as.dist(qiime.data$distmat) ~ qiime.data$map[[opts$category]],      permutations = opts$num_permutations) 

Permutation: free

Number of permutations: 999

Terms added sequentially (first to last)

                                Df SumsOfSqs MeanSqs F.Model      R2 Pr(>F)    

qiime.data$map[[opts$category]]  1   0.62181 0.62181  4.9462 0.27561  0.001 ***

Residuals                       13   1.63428 0.12571         0.72439           

Total                           14   2.25609                 1.00000           

---

Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

method name PERMANOVA

test statistic name pseudo-F

sample size 15

number of groups 4

test statistic 2.613314287290613

p-value 0.001

number of permutations 999

I am not sure that small sample size in redox (4 groups: 6 samples for one group and 3 samples each for the rest) is the problem. Treatment only has two groups: 6 samples for control and 9 samples for osa.

Could anyone can help me to clarify this problem, please? (attached file is my mapping file and distance matrix file).

Thank you very much!

Best regards,

An

[Jai Ram Rideout]

Hi An,

PERMANOVA in compare_categories.py treats any mapping file column as categorical data, even if it looks to be numeric/continuous. adonis in compare_categories.py runs the R vegan::adonis function under the hood, and this method performs a different test for continuous variables. Thus, you should see the same results between PERMANOVA and adonis when using categorical data (e.g. your "treatment" column), but you'll get different results from continuous data (e.g. your "redox condition" column).

Best,

Jai

[baoanxh2006]

Dear Jai,

Thank you very much!

That is beautiful. I checked it and it clear to me now.

Could you please help me with anther question?

I transferred the data from Qiime to R using Phyloseq. 

When I do PCoA and constrained analysis as well as Adonis, I cannot reproduce the same result as in Qiime: 

Such as a slight different in result for Adonis or a different variation explained by PC1/2 in PCoA.

I supposed something related to rarefying step. When I did not put rngseed=T, the results were different each time I run it. 

Please see detail in the attached file.

Thanks again for your great help!

Best regards,

An

[Jai Ram Rideout]

Hi An,

It's difficult to pinpoint why you're getting slightly different results between the two tools. Can you try using the same rarefied table in QIIME and phyloseq to produce PCoA plots, and compare those results?

Jai

[baoanxh2006]

Thanks Jai,

I will try and let you know how it going.

Best regards,

An